license: cc-by-sa-4.0
language:
- en
tags:
- regalfire
- evaluation
- reproducible
configs:
- config_name: default
data_files:
- split: train
path: train.jsonl
- split: validation
path: validation.jsonl
- split: test
path: test.jsonl
- split: holdout
path: holdout.jsonl
RAG-HardCases-MultiDomain
RegalFire — AI Data Foundry · Omar Soliman · ootiris@gmail.com
Problem
Multidomain scientific and fictional-policy RAG/evidence cases. 1,115 records in this source-derived release. Do not sum it with its public ancestors as unique examples.
Sources and provenance
SOURCE_LOCK.json pins approved releases and file hashes. Per-record source IDs and hashes link to source rows. Selected normalized native scientific-computing sources accompany the affected products in native_scicomp_source.jsonl; raw API scrape files are not included. Source URLs, author attribution and revision contributors are retained. Unverified Bionic inputs are excluded.
Schema and extraction
See schema.json for required fields/types. Nested source content and labels are independently validated. Scientific flags use acceptance, score order, answer count, literal citations and thread age. Multiple answers are not labelled conflicting/disagreeing. GCC answers are independently recomputed. Duplicate template/value content is removed.
What labels mean
Scientific records are metadata_grounded with no canonical truth; GCC records are deterministic_source_grounded with fictional answers and evidence.
Validation and reproduction
Install Python 3.10+ dependencies: pip install jsonschema huggingface_hub. Run in the repository directory: python validator/validate.py . Pinned public ancestors require network or a populated HF cache; credentials are unnecessary. The validator does not import the builder. Source computational checkers were implemented independently from source generators. See validation_report.json, robustness_report.json and QA_REPORT.md.
Splits and leakage
{"holdout": 192, "test": 199, "train": 653, "validation": 71}. Scientific threads, normalized shared text/code and near questions (5-shingle Jaccard >=0.85) are globally grouped across the two scientific-derived products. Synthetic task families are kept in one split, with repeated source grid groups united. Source master seeds identify an entire generation run, not independent per-task seeds. These are template-held-out splits, not fresh-generation-seed experiments. Uneven splits are intentional. Public holdout labels are visible and are not secret or contamination-free. Exhaustive semantic near-duplicate detection is not claimed.
Intended uses and limitations
Research, regression and scoped private adaptation. No measured model performance, clients, ROI or clinical claims. Scientific code is not executed; community answers are not expert labels. Clinical flags and email-like content are excluded heuristically, not certified absent. Agent errors are controlled harness operations, not real-agent logs; counterfactual recovery is prediction correction, not physical intervention. Validation checks the specified contracts, not universal scientific correctness.
License
cc-by-sa-4.0. Preserve attribution, revision contributors and source notices. Scientific content remains CC BY-SA 4.0; MIT synthetic rows retain MIT notices. See THIRD_PARTY_LICENSES.md. No private exclusivity is promised for public source adaptations.
Custom / Private Dataset Work
RegalFire builds custom AI datasets, evaluation sets and reproducible data pipelines for research and production systems.
Available:
- custom schemas
- private sources
- private holdout sets
- custom validators
- domain-specific hard cases
- continuous generation
- regression datasets
Contact: ootiris@gmail.com Hugging Face: RegalFire
Public license obligations remain applicable. Confidential private work requires client-owned or otherwise permitted sources.